At LEXIN (Center for Legal Informatics, Universitas Indonesia), my undergraduate thesis investigates how LLMs can translate Indonesian judicial reasoning into readable symbolic representations. The aim is to preserve what a selected passage says, including its claims, qualifications, attribution, and stated grounds.
I built a Python prototype for preparing court passages, generating schema-constrained model outputs, validating references, and applying bounded corrections. A separate model reconstructs Indonesian text using only the representation. Comparing the source, representation, and reconstruction helps inspect missing claims, unsupported additions, or changes to actors and conditions.
The experiments use versioned zero-shot and few-shot prompts and retain requests, outputs, validation reports, and correction artifacts so individual runs can be reviewed and repeated.
The research is ongoing. A structurally valid output or fluent reconstruction does not establish that the translation is faithful; source-conditioned review and broader semantic evaluation remain part of the work.